[DAGCombiner] (add (umax X, C), -C) --> (usubsat X, C) (PR40111)

[RKSimon]

Summary

Move the (add (umax X, C), -C) --> (usubsat X, C) X86 combine into generic DAGCombiner

First of a number of saturated arithmetic folds that can be moved out of X86-specific code for PR40111.

[nikic]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

2164 ↗

(On Diff #185307)

I'm not familiar with the nuances of the legality checking here, but wondering if using the hasOperation(ISD::USUBSAT, VT) helper would make more sense? If usubsat is not legal or custom, then the created usubsat is going to be expanded back into umax+add during legalization again anyway.

[RKSimon]

Use hasOperation

[nikic]

LGTM

llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
     SDValue N11 = N1.getOperand(1);
 
     if (isConstantOrConstantVector(N00) || isConstantOrConstantVector(N10))
       return DAG.getNode(ISD::SUB, DL, VT,
                          DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10),
                          DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11));
   }
 
+  // fold (add (umax X, C), -C) --> (usubsat X, C)
+  if (N0.getOpcode() == ISD::UMAX && hasOperation(ISD::USUBSAT, VT)) {
+    auto MatchUSUBSAT = [](ConstantSDNode *Max, ConstantSDNode *Op) {
+      return (!Max && !Op) ||
+             (Max && Op && Max->getAPIntValue() == (-Op->getAPIntValue()));
+    };
+    if (ISD::matchBinaryPredicate(N0.getOperand(1), N1, MatchUSUBSAT,
+                                  /*AllowUndefs*/ true))
+      return DAG.getNode(ISD::USUBSAT, DL, VT, N0.getOperand(0),
+                         N0.getOperand(1));
+  }
+
   if (SDValue V = foldAddSubBoolOfMaskedVal(N, DAG))
     return V;
 
   if (SDValue V = foldAddSubOfSignBit(N, DAG))
     return V;
 
   if (SimplifyDemandedBits(SDValue(N, 0)))
     return SDValue(N, 0);

llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

return DAG.getNode(X86ISD::VPMADDWD, DL, ResVT,
                        DAG.getNode(ISD::TRUNCATE, DL, TruncVT, Ops[0]),
                        DAG.getNode(ISD::TRUNCATE, DL, TruncVT, Ops[1]));
   };
   return SplitOpsAndApply(DAG, Subtarget, DL, VT,
                           { Mul.getOperand(0), Mul.getOperand(1) },
                           PMADDBuilder);
 }
 
-// Try to turn (add (umax X, C), -C) into (psubus X, C)
-static SDValue combineAddToSUBUS(SDNode *N, SelectionDAG &DAG,
-                                 const X86Subtarget &Subtarget) {
-  if (!Subtarget.hasSSE2())
-    return SDValue();
-
-  EVT VT = N->getValueType(0);
-
-  // psubus is available in SSE2 for i8 and i16 vectors.
-  if (!VT.isVector() || VT.getVectorNumElements() < 2 ||
-      !isPowerOf2_32(VT.getVectorNumElements()) ||
-      !(VT.getVectorElementType() == MVT::i8 ||
-        VT.getVectorElementType() == MVT::i16))
-    return SDValue();
-
-  SDValue Op0 = N->getOperand(0);
-  SDValue Op1 = N->getOperand(1);
-  if (Op0.getOpcode() != ISD::UMAX)
-    return SDValue();
-
-  // The add should have a constant that is the negative of the max.
-  auto MatchUSUBSAT = [](ConstantSDNode *Max, ConstantSDNode *Op) {
-    return (!Max && !Op) ||
-           (Max && Op && Max->getAPIntValue() == (-Op->getAPIntValue()));
-  };
-  if (!ISD::matchBinaryPredicate(Op0.getOperand(1), Op1, MatchUSUBSAT,
-                                 /*AllowUndefs*/ true))
-    return SDValue();
-
-  SDLoc DL(N);
-  return DAG.getNode(ISD::USUBSAT, DL, VT, Op0.getOperand(0),
-                     Op0.getOperand(1));
-}
-
 // Attempt to turn this pattern into PMADDWD.
 // (mul (add (zext (build_vector)), (zext (build_vector))),
 //      (add (zext (build_vector)), (zext (build_vector)))
 static SDValue matchPMADDWD_2(SelectionDAG &DAG, SDValue N0, SDValue N1,
                               const SDLoc &DL, EVT VT,
                               const X86Subtarget &Subtarget) {
   if (!Subtarget.hasSSE2())
     return SDValue();
if ((VT == MVT::v8i16 || VT == MVT::v4i32 || VT == MVT::v16i16 ||
     };
     return SplitOpsAndApply(DAG, Subtarget, SDLoc(N), VT, {Op0, Op1},
                             HADDBuilder);
   }
 
   if (SDValue V = combineIncDecVector(N, DAG))
     return V;
 
-  if (SDValue V = combineAddToSUBUS(N, DAG, Subtarget))
-    return V;
-
   return combineAddOrSubToADCOrSBB(N, DAG);
 }
 
 static SDValue combineSubToSubus(SDNode *N, SelectionDAG &DAG,
                                  const X86Subtarget &Subtarget) {
   SDValue Op0 = N->getOperand(0);
   SDValue Op1 = N->getOperand(1);
   EVT VT = N->getValueType(0);